lkml.org 
[lkml]   [2017]   [Nov]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
SubjectRe: [RFC -mm] mm, userfaultfd, THP: Avoid waiting when PMD under THP migration
Date
On 4 Nov 2017, at 23:01, huang ying wrote:

> On Fri, Nov 3, 2017 at 11:00 PM, Zi Yan <zi.yan@cs.rutgers.edu> wrote:
>> On 3 Nov 2017, at 3:52, Huang, Ying wrote:
>>
>>> From: Huang Ying <ying.huang@intel.com>
>>>
>>> If THP migration is enabled, the following situation is possible,
>>>
>>> - A THP is mapped at source address
>>> - Migration is started to move the THP to another node
>>> - Page fault occurs
>>> - The PMD (migration entry) is copied to the destination address in
>>> mremap
>>>
>>
>> You mean the page fault path follows the source address and sees
>> pmd_none() now
>> because mremap() clears it and remaps the page with dest address.
>> Otherwise, it seems not possible to get into handle_userfault(),
>> since it is called in
>> pmd_none() branch inside do_huge_pmd_anonymous_page().
>>
>>
>>> That is, it is possible for handle_userfault() encounter a PMD entry
>>> which has been handled but !pmd_present(). In the current
>>> implementation, we will wait for such PMD entries, which may cause
>>> unnecessary waiting, and potential soft lockup.
>>
>> handle_userfault() should only see pmd_none() in the situation you
>> describe,
>> whereas !pmd_present() (migration entry case) should lead to
>> pmd_migration_entry_wait().
>
> Yes. This is my understanding of the source code too. And I
> described it in the original patch description too. I just want to
> make sure whether it is possible that !pmd_none() and !pmd_present()
> for a PMD in userfaultfd_must_wait(). And, whether it is possible for
> us to implement PMD mapping copying in UFFDIO_COPY in the future?
>

Thanks for clarifying it. We both agree that !pmd_present(), which means
PMD migration entry, does not get into userfaultfd_must_wait(),
then there seems to be no issue with current code yet.

However, the if (!pmd_present(_pmd)) in userfaultfd_must_wait() does not
match
the exact condition. How about the patch below? It can catch pmd
migration entries,
which are only possible in x86_64 at the moment.

diff --git a/fs/userfaultfd.c b/fs/userfaultfd.c
index 1c713fd5b3e6..dda25444a6ee 100644
--- a/fs/userfaultfd.c
+++ b/fs/userfaultfd.c
@@ -294,9 +294,11 @@ static inline bool userfaultfd_must_wait(struct
userfaultfd_ctx *ctx,
* pmd_trans_unstable) of the pmd.
*/
_pmd = READ_ONCE(*pmd);
- if (!pmd_present(_pmd))
+ if (pmd_none(_pmd))
goto out;

+ VM_BUG_ON(thp_migration_supported() &&
is_pmd_migration_entry(_pmd));
+
ret = false;
if (pmd_trans_huge(_pmd))
goto out;



Best Regards,
Yan Zi

\
 
 \ /
  Last update: 2017-11-06 16:54    [W:0.079 / U:1.844 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site